Item: APPLYING THE AVALANCHE DANGER SCALE: CONSISTENCY DEPENDS MORE ON THE SITUATION THAN THE FORECASTER
-
-
Title: APPLYING THE AVALANCHE DANGER SCALE: CONSISTENCY DEPENDS MORE ON THE SITUATION THAN THE FORECASTER
Proceedings: Proceedings, International Snow Science Workshop, Whistler, BC, Canada, 2026
Authors:
- Brian Lazar [ Colorado Avalanche Information Center, Denver, CO USA ]
- Dylan Craaybeek [ Colorado Avalanche Information Center, Denver, CO USA ]
- Frank Techel [ WSL Institute for Snow and Avalanche Research SLF, Davos, Switzerland ]
- Simon Horton [ Avalanche Canada, Revelstoke, BC, Canada ]
- Mike Cooperstein [ Colorado Avalanche Information Center, Denver, CO USA ]
- Simon Trautman [ American Avalanche Association, Jackson, WY USA ]
- Ethan Greene [ Colorado Avalanche Information Center, Denver, CO USA ]
Date: 2026-09-28
Abstract: Public avalanche forecasts compress complex hazard assessments into a single danger level to guide backcountry recreationists' terrain choices. Danger increases exponentially between levels, so a one-level difference carries significant behavioral and public safety consequences. A decade after Lazar et al. (2016) demonstrated substantial forecaster disagreement, we repeated the exercise with 144 professional public avalanche forecasters across North America, Europe, and New Zealand, each evaluating ten simulated mountain-range-scale scenarios. Two randomly selected forecasters agreed on the exact danger level only 54% of the time (Krippendorff's ordinal α = 0.65, 95% CI: 0.33 to 0.78), though 96% remained within one level. Agreement did not differ clearly by region or experience. Instead, consistency was dominated by scenario complexity (exact agreement ranged from 41% to 75%). In contrast, agreement on the primary avalanche problem was substantially higher, exceeding 80% consensus on half the scenarios, suggesting that disagreement may stem from danger-scale threshold decisions rather than hazard identification. When respondents deviated from the reference rating, they compressed toward the middle of the scale, over-forecasting low-hazard conditions and under-forecasting high-hazard conditions. Pooling independent forecasters and utilizing the group median significantly raised inter-panel reproducibility. A panel of three increased agreement by 13 percentage points (to 67%), with the first two added forecasters improving agreement more than six times as much per forecaster as subsequent additions. However, group aggregation only reduces random individual noise, leaving shared systematic biases intact. We recommend drafting your ratings independently first, and then discuss them as a group. If you are forecasting solo, an external opinion is valuable, whether from a neighboring office or a model, because a team cannot detect a bias its members share.
Object ID: ISSW2026_O9.3.pdf
DOI: https://doi.org/10.15788/1790098689
Language of Article: English
Presenter(s): Brian Lazar
Keywords: avalanche danger scale, forecaster consistency, inter-rater reliability, judgment aggregation, avalanche problems, public avalanche forecasting
Page Number(s): 535 - 542
-